Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,268 +1,268 @@
|
|
1 |
-
import glob
|
2 |
-
import os
|
3 |
-
import shutil
|
4 |
-
import sys
|
5 |
-
import re
|
6 |
-
import tempfile
|
7 |
-
import zipfile
|
8 |
-
from pathlib import Path
|
9 |
-
|
10 |
-
import gradio as gr
|
11 |
-
|
12 |
-
from finetune import finetune_model, baseline_model
|
13 |
-
|
14 |
-
from language import languages
|
15 |
-
from task import tasks
|
16 |
-
import matplotlib.pyplot as plt
|
17 |
-
|
18 |
-
|
19 |
-
os.environ['TEMP_DIR'] = tempfile.mkdtemp()
|
20 |
-
|
21 |
-
def load_markdown():
|
22 |
-
with open("intro.md", "r") as f:
|
23 |
-
return f.read()
|
24 |
-
|
25 |
-
|
26 |
-
def read_logs():
|
27 |
-
try:
|
28 |
-
with open(f"output.log", "r") as f:
|
29 |
-
return f.read()
|
30 |
-
except:
|
31 |
-
return None
|
32 |
-
|
33 |
-
|
34 |
-
def plot_loss_acc(temp_dir, log_every):
|
35 |
-
sys.stdout.flush()
|
36 |
-
lines = []
|
37 |
-
with open("output.log", "r") as f:
|
38 |
-
for line in f.readlines():
|
39 |
-
if re.match(r"^\[\d+\] - loss: \d+\.\d+ - acc: \d+\.\d+$", line):
|
40 |
-
lines.append(line)
|
41 |
-
|
42 |
-
losses = []
|
43 |
-
acces = []
|
44 |
-
if len(lines) == 0:
|
45 |
-
return None, None
|
46 |
-
|
47 |
-
for line in lines:
|
48 |
-
_, loss, acc = line.split(" - ")
|
49 |
-
losses.append(float(loss.split(":")[1].strip()))
|
50 |
-
acces.append(float(acc.split(":")[1].strip()))
|
51 |
-
|
52 |
-
x = [i * log_every for i in range(1, len(losses) + 1)]
|
53 |
-
|
54 |
-
plt.plot(x, losses, label="loss")
|
55 |
-
plt.xlim(log_every // 2, x[-1] + log_every // 2)
|
56 |
-
plt.savefig(f"{temp_dir}/loss.png")
|
57 |
-
plt.clf()
|
58 |
-
plt.plot(x, acces, label="acc")
|
59 |
-
plt.xlim(log_every // 2, x[-1] + log_every // 2)
|
60 |
-
plt.savefig(f"{temp_dir}/acc.png")
|
61 |
-
plt.clf()
|
62 |
-
return f"{temp_dir}/acc.png", f"{temp_dir}/loss.png"
|
63 |
-
|
64 |
-
|
65 |
-
def upload_file(fileobj, temp_dir):
|
66 |
-
"""
|
67 |
-
Upload a file and check the uploaded zip file.
|
68 |
-
"""
|
69 |
-
# First check if a file is a zip file.
|
70 |
-
if not zipfile.is_zipfile(fileobj.name):
|
71 |
-
raise gr.Error("Please upload a zip file.")
|
72 |
-
|
73 |
-
# Then unzip file
|
74 |
-
shutil.unpack_archive(fileobj.name, temp_dir)
|
75 |
-
|
76 |
-
# check zip file
|
77 |
-
if not os.path.exists(os.path.join(temp_dir, "text")):
|
78 |
-
raise gr.Error("Please upload a valid zip file.")
|
79 |
-
|
80 |
-
if not os.path.exists(os.path.join(temp_dir, "text_ctc")):
|
81 |
-
raise gr.Error("Please upload a valid zip file.")
|
82 |
-
|
83 |
-
if not os.path.exists(os.path.join(temp_dir, "audio")):
|
84 |
-
raise gr.Error("Please upload a valid zip file.")
|
85 |
-
|
86 |
-
# check if all texts and audio matches
|
87 |
-
audio_ids = []
|
88 |
-
with open(os.path.join(temp_dir, "text"), "r") as f:
|
89 |
-
for line in f.readlines():
|
90 |
-
audio_ids.append(line.split(maxsplit=1)[0])
|
91 |
-
|
92 |
-
with open(os.path.join(temp_dir, "text_ctc"), "r") as f:
|
93 |
-
ctc_audio_ids = []
|
94 |
-
for line in f.readlines():
|
95 |
-
ctc_audio_ids.append(line.split(maxsplit=1)[0])
|
96 |
-
|
97 |
-
if len(audio_ids) != len(ctc_audio_ids):
|
98 |
-
raise gr.Error(
|
99 |
-
f"Length of `text` ({len(audio_ids)}) and `text_ctc` ({len(ctc_audio_ids)}) is different."
|
100 |
-
)
|
101 |
-
|
102 |
-
if set(audio_ids) != set(ctc_audio_ids):
|
103 |
-
raise gr.Error(f"`text` and `text_ctc` have different audio ids.")
|
104 |
-
|
105 |
-
for audio_id in glob.glob(os.path.join(temp_dir, "audio", "*")):
|
106 |
-
if not Path(audio_id).stem in audio_ids:
|
107 |
-
raise gr.Error(f"Audio id {audio_id} is not in `text` or `text_ctc`.")
|
108 |
-
|
109 |
-
gr.Info("Successfully uploaded and validated zip file.")
|
110 |
-
|
111 |
-
return [fileobj]
|
112 |
-
|
113 |
-
|
114 |
-
with gr.Blocks(title="OWSM-finetune") as demo:
|
115 |
-
tempdir_path = gr.State(os.environ['TEMP_DIR'])
|
116 |
-
gr.Markdown(
|
117 |
-
"""# OWSM finetune demo!
|
118 |
-
|
119 |
-
Finetune `owsm_v3.1_ebf_base` with your own dataset!
|
120 |
-
Due to resource limitation, you can only train
|
121 |
-
|
122 |
-
## Upload dataset and define settings
|
123 |
-
"""
|
124 |
-
)
|
125 |
-
|
126 |
-
# main contents
|
127 |
-
with gr.Row():
|
128 |
-
with gr.Column():
|
129 |
-
file_output = gr.File()
|
130 |
-
upload_button = gr.UploadButton("Click to Upload a File", file_count="single")
|
131 |
-
upload_button.upload(
|
132 |
-
upload_file, [upload_button, tempdir_path], [file_output]
|
133 |
-
)
|
134 |
-
|
135 |
-
with gr.Column():
|
136 |
-
lang = gr.Dropdown(
|
137 |
-
languages["espnet/owsm_v3.1_ebf_base"],
|
138 |
-
label="Language",
|
139 |
-
info="Choose language!",
|
140 |
-
value="jpn",
|
141 |
-
interactive=True,
|
142 |
-
)
|
143 |
-
task = gr.Dropdown(
|
144 |
-
tasks["espnet/owsm_v3.1_ebf_base"],
|
145 |
-
label="Task",
|
146 |
-
info="Choose task!",
|
147 |
-
value="asr",
|
148 |
-
interactive=True,
|
149 |
-
)
|
150 |
-
|
151 |
-
gr.Markdown("## Set training settings")
|
152 |
-
|
153 |
-
with gr.Row():
|
154 |
-
with gr.Column():
|
155 |
-
log_every = gr.Number(value=10, label="log_every", interactive=True)
|
156 |
-
max_epoch = gr.Slider(1, 10, step=1, label="max_epoch", interactive=True)
|
157 |
-
scheduler = gr.Dropdown(
|
158 |
-
["warmuplr"], label="warmup", value="warmuplr", interactive=True
|
159 |
-
)
|
160 |
-
warmup_steps = gr.Number(
|
161 |
-
value=100, label="warmup_steps", interactive=True
|
162 |
-
)
|
163 |
-
|
164 |
-
with gr.Column():
|
165 |
-
optimizer = gr.Dropdown(
|
166 |
-
["adam", "adamw", "sgd", "adadelta", "adagrad", "adamax", "asgd", "rmsprop"],
|
167 |
-
label="optimizer",
|
168 |
-
value="adam",
|
169 |
-
interactive=True
|
170 |
-
)
|
171 |
-
learning_rate = gr.Number(
|
172 |
-
value=1e-4, label="learning_rate", interactive=True
|
173 |
-
)
|
174 |
-
weight_decay = gr.Number(
|
175 |
-
value=0.000001, label="weight_decay", interactive=True
|
176 |
-
)
|
177 |
-
|
178 |
-
gr.Markdown("## Logs and plots")
|
179 |
-
|
180 |
-
with gr.Row():
|
181 |
-
with gr.Column():
|
182 |
-
log_output = gr.Textbox(
|
183 |
-
show_label=False,
|
184 |
-
interactive=False,
|
185 |
-
max_lines=23,
|
186 |
-
lines=23,
|
187 |
-
)
|
188 |
-
demo.load(read_logs, None, log_output, every=2)
|
189 |
-
|
190 |
-
with gr.Column():
|
191 |
-
log_acc = gr.Image(label="Accuracy", show_label=True, interactive=False)
|
192 |
-
log_loss = gr.Image(label="Loss", show_label=True, interactive=False)
|
193 |
-
demo.load(plot_loss_acc, [tempdir_path, log_every], [log_acc, log_loss], every=10)
|
194 |
-
|
195 |
-
with gr.Row():
|
196 |
-
with gr.Column():
|
197 |
-
ref_text = gr.Textbox(
|
198 |
-
label="Reference text",
|
199 |
-
show_label=True,
|
200 |
-
interactive=False,
|
201 |
-
max_lines=10,
|
202 |
-
lines=10,
|
203 |
-
)
|
204 |
-
with gr.Column():
|
205 |
-
base_text = gr.Textbox(
|
206 |
-
label="Baseline text",
|
207 |
-
show_label=True,
|
208 |
-
interactive=False,
|
209 |
-
max_lines=10,
|
210 |
-
lines=10,
|
211 |
-
)
|
212 |
-
|
213 |
-
with gr.Row():
|
214 |
-
with gr.Column():
|
215 |
-
hyp_text = gr.Textbox(
|
216 |
-
label="Hypothesis text",
|
217 |
-
show_label=True,
|
218 |
-
interactive=False,
|
219 |
-
max_lines=10,
|
220 |
-
lines=10,
|
221 |
-
)
|
222 |
-
with gr.Column():
|
223 |
-
trained_model = gr.File(
|
224 |
-
label="Trained model",
|
225 |
-
interactive=False,
|
226 |
-
)
|
227 |
-
|
228 |
-
with gr.Row():
|
229 |
-
with gr.Column():
|
230 |
-
baseline_btn = gr.Button("Run Baseline", variant="secondary")
|
231 |
-
baseline_btn.click(
|
232 |
-
baseline_model,
|
233 |
-
[
|
234 |
-
lang,
|
235 |
-
task,
|
236 |
-
tempdir_path,
|
237 |
-
],
|
238 |
-
[ref_text, base_text]
|
239 |
-
)
|
240 |
-
with gr.Column():
|
241 |
-
finetune_btn = gr.Button("Finetune Model", variant="primary")
|
242 |
-
finetune_btn.click(
|
243 |
-
finetune_model,
|
244 |
-
[
|
245 |
-
lang,
|
246 |
-
task,
|
247 |
-
tempdir_path,
|
248 |
-
log_every,
|
249 |
-
max_epoch,
|
250 |
-
scheduler,
|
251 |
-
warmup_steps,
|
252 |
-
optimizer,
|
253 |
-
learning_rate,
|
254 |
-
weight_decay,
|
255 |
-
],
|
256 |
-
[trained_model, hyp_text]
|
257 |
-
)
|
258 |
-
|
259 |
-
gr.Markdown(load_markdown())
|
260 |
-
|
261 |
-
if __name__ == "__main__":
|
262 |
-
try:
|
263 |
-
demo.queue().launch()
|
264 |
-
except:
|
265 |
-
print("Unexpected error:", sys.exc_info()[0])
|
266 |
-
raise
|
267 |
-
finally:
|
268 |
-
shutil.rmtree(os.environ['TEMP_DIR'])
|
|
|
1 |
+
import glob
|
2 |
+
import os
|
3 |
+
import shutil
|
4 |
+
import sys
|
5 |
+
import re
|
6 |
+
import tempfile
|
7 |
+
import zipfile
|
8 |
+
from pathlib import Path
|
9 |
+
|
10 |
+
import gradio as gr
|
11 |
+
|
12 |
+
from finetune import finetune_model, baseline_model
|
13 |
+
|
14 |
+
from language import languages
|
15 |
+
from task import tasks
|
16 |
+
import matplotlib.pyplot as plt
|
17 |
+
|
18 |
+
|
19 |
+
os.environ['TEMP_DIR'] = tempfile.mkdtemp()
|
20 |
+
|
21 |
+
def load_markdown():
|
22 |
+
with open("intro.md", "r") as f:
|
23 |
+
return f.read()
|
24 |
+
|
25 |
+
|
26 |
+
def read_logs():
|
27 |
+
try:
|
28 |
+
with open(f"output.log", "r") as f:
|
29 |
+
return f.read()
|
30 |
+
except:
|
31 |
+
return None
|
32 |
+
|
33 |
+
|
34 |
+
def plot_loss_acc(temp_dir, log_every):
|
35 |
+
sys.stdout.flush()
|
36 |
+
lines = []
|
37 |
+
with open("output.log", "r") as f:
|
38 |
+
for line in f.readlines():
|
39 |
+
if re.match(r"^\[\d+\] - loss: \d+\.\d+ - acc: \d+\.\d+$", line):
|
40 |
+
lines.append(line)
|
41 |
+
|
42 |
+
losses = []
|
43 |
+
acces = []
|
44 |
+
if len(lines) == 0:
|
45 |
+
return None, None
|
46 |
+
|
47 |
+
for line in lines:
|
48 |
+
_, loss, acc = line.split(" - ")
|
49 |
+
losses.append(float(loss.split(":")[1].strip()))
|
50 |
+
acces.append(float(acc.split(":")[1].strip()))
|
51 |
+
|
52 |
+
x = [i * log_every for i in range(1, len(losses) + 1)]
|
53 |
+
|
54 |
+
plt.plot(x, losses, label="loss")
|
55 |
+
plt.xlim(log_every // 2, x[-1] + log_every // 2)
|
56 |
+
plt.savefig(f"{temp_dir}/loss.png")
|
57 |
+
plt.clf()
|
58 |
+
plt.plot(x, acces, label="acc")
|
59 |
+
plt.xlim(log_every // 2, x[-1] + log_every // 2)
|
60 |
+
plt.savefig(f"{temp_dir}/acc.png")
|
61 |
+
plt.clf()
|
62 |
+
return f"{temp_dir}/acc.png", f"{temp_dir}/loss.png"
|
63 |
+
|
64 |
+
|
65 |
+
def upload_file(fileobj, temp_dir):
|
66 |
+
"""
|
67 |
+
Upload a file and check the uploaded zip file.
|
68 |
+
"""
|
69 |
+
# First check if a file is a zip file.
|
70 |
+
if not zipfile.is_zipfile(fileobj.name):
|
71 |
+
raise gr.Error("Please upload a zip file.")
|
72 |
+
|
73 |
+
# Then unzip file
|
74 |
+
shutil.unpack_archive(fileobj.name, temp_dir)
|
75 |
+
|
76 |
+
# check zip file
|
77 |
+
if not os.path.exists(os.path.join(temp_dir, "text")):
|
78 |
+
raise gr.Error("Please upload a valid zip file.")
|
79 |
+
|
80 |
+
if not os.path.exists(os.path.join(temp_dir, "text_ctc")):
|
81 |
+
raise gr.Error("Please upload a valid zip file.")
|
82 |
+
|
83 |
+
if not os.path.exists(os.path.join(temp_dir, "audio")):
|
84 |
+
raise gr.Error("Please upload a valid zip file.")
|
85 |
+
|
86 |
+
# check if all texts and audio matches
|
87 |
+
audio_ids = []
|
88 |
+
with open(os.path.join(temp_dir, "text"), "r") as f:
|
89 |
+
for line in f.readlines():
|
90 |
+
audio_ids.append(line.split(maxsplit=1)[0])
|
91 |
+
|
92 |
+
with open(os.path.join(temp_dir, "text_ctc"), "r") as f:
|
93 |
+
ctc_audio_ids = []
|
94 |
+
for line in f.readlines():
|
95 |
+
ctc_audio_ids.append(line.split(maxsplit=1)[0])
|
96 |
+
|
97 |
+
if len(audio_ids) != len(ctc_audio_ids):
|
98 |
+
raise gr.Error(
|
99 |
+
f"Length of `text` ({len(audio_ids)}) and `text_ctc` ({len(ctc_audio_ids)}) is different."
|
100 |
+
)
|
101 |
+
|
102 |
+
if set(audio_ids) != set(ctc_audio_ids):
|
103 |
+
raise gr.Error(f"`text` and `text_ctc` have different audio ids.")
|
104 |
+
|
105 |
+
for audio_id in glob.glob(os.path.join(temp_dir, "audio", "*")):
|
106 |
+
if not Path(audio_id).stem in audio_ids:
|
107 |
+
raise gr.Error(f"Audio id {audio_id} is not in `text` or `text_ctc`.")
|
108 |
+
|
109 |
+
gr.Info("Successfully uploaded and validated zip file.")
|
110 |
+
|
111 |
+
return [fileobj]
|
112 |
+
|
113 |
+
|
114 |
+
with gr.Blocks(title="OWSM-finetune") as demo:
|
115 |
+
tempdir_path = gr.State(os.environ['TEMP_DIR'])
|
116 |
+
gr.Markdown(
|
117 |
+
"""# OWSM finetune demo!
|
118 |
+
|
119 |
+
Finetune `owsm_v3.1_ebf_base` with your own dataset!
|
120 |
+
Due to resource limitation, you can only train 10 epochs on maximum.
|
121 |
+
|
122 |
+
## Upload dataset and define settings
|
123 |
+
"""
|
124 |
+
)
|
125 |
+
|
126 |
+
# main contents
|
127 |
+
with gr.Row():
|
128 |
+
with gr.Column():
|
129 |
+
file_output = gr.File()
|
130 |
+
upload_button = gr.UploadButton("Click to Upload a File", file_count="single")
|
131 |
+
upload_button.upload(
|
132 |
+
upload_file, [upload_button, tempdir_path], [file_output]
|
133 |
+
)
|
134 |
+
|
135 |
+
with gr.Column():
|
136 |
+
lang = gr.Dropdown(
|
137 |
+
languages["espnet/owsm_v3.1_ebf_base"],
|
138 |
+
label="Language",
|
139 |
+
info="Choose language!",
|
140 |
+
value="jpn",
|
141 |
+
interactive=True,
|
142 |
+
)
|
143 |
+
task = gr.Dropdown(
|
144 |
+
tasks["espnet/owsm_v3.1_ebf_base"],
|
145 |
+
label="Task",
|
146 |
+
info="Choose task!",
|
147 |
+
value="asr",
|
148 |
+
interactive=True,
|
149 |
+
)
|
150 |
+
|
151 |
+
gr.Markdown("## Set training settings")
|
152 |
+
|
153 |
+
with gr.Row():
|
154 |
+
with gr.Column():
|
155 |
+
log_every = gr.Number(value=10, label="log_every", interactive=True)
|
156 |
+
max_epoch = gr.Slider(1, 10, step=1, label="max_epoch", interactive=True)
|
157 |
+
scheduler = gr.Dropdown(
|
158 |
+
["warmuplr"], label="warmup", value="warmuplr", interactive=True
|
159 |
+
)
|
160 |
+
warmup_steps = gr.Number(
|
161 |
+
value=100, label="warmup_steps", interactive=True
|
162 |
+
)
|
163 |
+
|
164 |
+
with gr.Column():
|
165 |
+
optimizer = gr.Dropdown(
|
166 |
+
["adam", "adamw", "sgd", "adadelta", "adagrad", "adamax", "asgd", "rmsprop"],
|
167 |
+
label="optimizer",
|
168 |
+
value="adam",
|
169 |
+
interactive=True
|
170 |
+
)
|
171 |
+
learning_rate = gr.Number(
|
172 |
+
value=1e-4, label="learning_rate", interactive=True
|
173 |
+
)
|
174 |
+
weight_decay = gr.Number(
|
175 |
+
value=0.000001, label="weight_decay", interactive=True
|
176 |
+
)
|
177 |
+
|
178 |
+
gr.Markdown("## Logs and plots")
|
179 |
+
|
180 |
+
with gr.Row():
|
181 |
+
with gr.Column():
|
182 |
+
log_output = gr.Textbox(
|
183 |
+
show_label=False,
|
184 |
+
interactive=False,
|
185 |
+
max_lines=23,
|
186 |
+
lines=23,
|
187 |
+
)
|
188 |
+
demo.load(read_logs, None, log_output, every=2)
|
189 |
+
|
190 |
+
with gr.Column():
|
191 |
+
log_acc = gr.Image(label="Accuracy", show_label=True, interactive=False)
|
192 |
+
log_loss = gr.Image(label="Loss", show_label=True, interactive=False)
|
193 |
+
demo.load(plot_loss_acc, [tempdir_path, log_every], [log_acc, log_loss], every=10)
|
194 |
+
|
195 |
+
with gr.Row():
|
196 |
+
with gr.Column():
|
197 |
+
ref_text = gr.Textbox(
|
198 |
+
label="Reference text",
|
199 |
+
show_label=True,
|
200 |
+
interactive=False,
|
201 |
+
max_lines=10,
|
202 |
+
lines=10,
|
203 |
+
)
|
204 |
+
with gr.Column():
|
205 |
+
base_text = gr.Textbox(
|
206 |
+
label="Baseline text",
|
207 |
+
show_label=True,
|
208 |
+
interactive=False,
|
209 |
+
max_lines=10,
|
210 |
+
lines=10,
|
211 |
+
)
|
212 |
+
|
213 |
+
with gr.Row():
|
214 |
+
with gr.Column():
|
215 |
+
hyp_text = gr.Textbox(
|
216 |
+
label="Hypothesis text",
|
217 |
+
show_label=True,
|
218 |
+
interactive=False,
|
219 |
+
max_lines=10,
|
220 |
+
lines=10,
|
221 |
+
)
|
222 |
+
with gr.Column():
|
223 |
+
trained_model = gr.File(
|
224 |
+
label="Trained model",
|
225 |
+
interactive=False,
|
226 |
+
)
|
227 |
+
|
228 |
+
with gr.Row():
|
229 |
+
with gr.Column():
|
230 |
+
baseline_btn = gr.Button("Run Baseline", variant="secondary")
|
231 |
+
baseline_btn.click(
|
232 |
+
baseline_model,
|
233 |
+
[
|
234 |
+
lang,
|
235 |
+
task,
|
236 |
+
tempdir_path,
|
237 |
+
],
|
238 |
+
[ref_text, base_text]
|
239 |
+
)
|
240 |
+
with gr.Column():
|
241 |
+
finetune_btn = gr.Button("Finetune Model", variant="primary")
|
242 |
+
finetune_btn.click(
|
243 |
+
finetune_model,
|
244 |
+
[
|
245 |
+
lang,
|
246 |
+
task,
|
247 |
+
tempdir_path,
|
248 |
+
log_every,
|
249 |
+
max_epoch,
|
250 |
+
scheduler,
|
251 |
+
warmup_steps,
|
252 |
+
optimizer,
|
253 |
+
learning_rate,
|
254 |
+
weight_decay,
|
255 |
+
],
|
256 |
+
[trained_model, hyp_text]
|
257 |
+
)
|
258 |
+
|
259 |
+
gr.Markdown(load_markdown())
|
260 |
+
|
261 |
+
if __name__ == "__main__":
|
262 |
+
try:
|
263 |
+
demo.queue().launch()
|
264 |
+
except:
|
265 |
+
print("Unexpected error:", sys.exc_info()[0])
|
266 |
+
raise
|
267 |
+
finally:
|
268 |
+
shutil.rmtree(os.environ['TEMP_DIR'])
|